What Data for This Category Looks Like
Typical users are process and quality departments at foundries and die casting workshops. Daily documents include foundry process cards, batching lists, heat records, mold history logs, first article inspection records, material test reports, and non-conforming product review forms. Data sources include ERP material master data, melting furnaces and thermometers, spectrometers, MES heat records, mold management systems, customer drawings, and technical requirements. Data updates occur per heat, per batch, or daily. Fields include alloy batching parameters, furnace temperature values, pouring temperature, mold usage counts, defect type records, and more. The data exists in both structured form and semi-structured document formats.
Constraints Imposed on Vector Models and Indexing
High-frequency heat and batch-level updates require indexes to support fast incremental synchronization, avoiding the time costs of full indexing. Semi-structured document formats require chunking strategies that balance parameter completeness and semantic coherence. This prevents critical process fields such as batching ratios and temperature values from being split across chunks. Unlike forging categories that rely on plastic deformation, foundry data centers around melting and pouring parameters for liquid metal forming. Indexes must therefore prioritize numerical fields such as temperature and batching ratios, adapting to encoding logic for numerical features to improve parameter matching accuracy. Process Q&A also requires linking traceability information such as heat and mold run counts. Indexes must retain corresponding metadata fields to ensure recall results can be traced back to specific production batches.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the single-paragraph process description length of foundry process documents, covering complete process steps and parameter combinations |
| Overlap Length | 100–200 characters | Prevents key parameters such as furnace temperature and batching ratios from being cut off during chunking, ensuring semantic coherence |
| Similarity Threshold | 0.75–0.85 | Filters low-match non-process-related documents, ensuring recalled content focuses on foundry process specifications |
| Single Search Limit | Top 10 results | Covers multiple types of associated documents involved in a single heat process, avoiding redundant results |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
| Reranking | Enabled | Performs secondary sorting on recall results by semantic relevance, improving the ability to accurately match process questions |
Three Common Misconfigurations
- Symptom: Incremental index update delay, with several hours of synchronization time after heat data is uploaded. Cause: Incremental indexing triggered by production batches is not configured, and full index updates are used instead. This fails to adapt to high-frequency heat data synchronization requirements.
- Symptom: Search results include unrelated non-foundry process documents. Cause: The similarity threshold is not adjusted for foundry category numerical parameters, leading to recall of cross-category process content.
- Symptom: Matching results for alloy batching ratios are missing. Cause: Chunk length is set too short, cutting off complete ratio parameters from batching lists. This prevents vector encoding from covering full process information.
How to Confirm Correct Configuration
- Upload a latest heat record document, check if the indexing task status in FastGPT v4.17.0 shows successful.
- Enter a query containing alloy batching parameters, verify that recalled results include the corresponding foundry process card or batching list document.
- Modify a mold history log, trigger an incremental update, then query related questions again to confirm recalled content has been synchronously updated.
- View the metadata list in the vector store, confirm that each index is associated with the corresponding heat and mold run identification fields.
The values provided are common starting points and should be evaluated against the reader's own samples.